Predicting the shear strength of FRP-RC beams using optimized CatBoost machine learning model

Abstract

Although FRP is an effective alternative to traditional steel reinforcement, determining the shear strength of FRP-RC beams remains a challenge due to the distinct properties of this material compared to steel. This study focuses on the development a database from collected experimental results and combining it with numerical simulations to develop a machine learning (ML) model capable of accurately predicting the shear strength of FRP-RC beams. This study utilizes data from 453 experimental results collected from 54 papers and combines it with numerical simulations to build a ML model based on CatBoost algorithm optimized by grid search and random search techniques. The optimal CatBoost model achieved high accuracy with R² values of 0.998 and 0.953, RMSE values of 5.647 and 21.908, corresponding to the training and testing datasets, respectively. The analysis of the influence of parameters on the shear strength of FRP-RC beams was conducted using the SHAP analysis technique, indicating that the effective depth, shear span-to-depth ratio, FRP stirrup ratio, and FRP longitudinal reinforcement ratio are the most important factors. This model not only provides a fast and efficient calculation tool but also helps quantify the influence of various factors on shear strength, thereby supporting the optimization of design and use of FRP materials in construction.